If you asked a model like ChatGPT or Perplexity which product in your category it would recommend yesterday, the answer order is almost certainly not what your SEO dashboard predicts. Half of the names come out of a surface your team has never been asked to maintain. The brands you expect to see at the top do not show, and the ones that do are often cited through sources you did not know existed. That is the budgeting question I keep seeing on the agenda for Q1 2027, and it begins with a simpler one: where is the model actually learning this about your brand.
The honest answer is that it depends on your category, and most “AI SEO” playbooks currently in circulation obscure that fact. Reddit is huge in beauty. Clinical and health publishers run supplements. For B2B software, review platforms and vertical directories are back, and one of them is often a vertical directory you have not thought about since 2019. Pointing a 2027 budget at Reddit because Reddit drives roughly 24% of Perplexity citations is as useful as pointing a 2015 budget at Facebook because people spend time there.
Three reports have landed in the last five months with real figures behind them, and they disagree with the generic AI-source charts in a very usable way. I want to walk through what each one actually shows, and then, as the agency that gets handed the ticket once the direction is set, spell out what the finding becomes inside a WordPress or WooCommerce estate on Monday morning.
The Universal Top-Source List Is Hiding Your Actual Discovery Layer
Every quarter now there is a new ranking of the sources large language models cite most. The headlines come out roughly the same: YouTube, Reddit, LinkedIn, Wikipedia, Forbes. Contently’s April 2026 synthesis across five independent studies puts that same five near the top. If you read only that chart you would conclude the recipe for AI visibility is to buy video, post on LinkedIn, and seed Reddit.
That reading is wrong, and the data under the chart is where it gets interesting. The same syntheses note that Reddit, Wikipedia, LinkedIn and YouTube combined rarely top 5% of total citations in any individual industry, and the other 95% spreads across thousands of domains. Which means the general chart tells you where to go hunting for crumbs, not where the loaf is.
The loaf, in every piece of client work I have reviewed this year, is category-specific. The sources that teach a model about a brand are the publishers, directories, review sites and marketplaces that your industry’s answers get drawn from. That is a different list for beauty, for supplements, for B2B software, for financial services. Three studies published since May put real figures to that split. I want to look at them in turn, because the shape of each one tells you what you are actually buying when you line-item AI discoverability on next year’s budget.
In Beauty, Nearly Two Thirds of AI Citations Come From Editorial Titles
Avenue Z published its 2026 AI Visibility Index for beauty on 20 July 2026, covering 60 brands and 565 ChatGPT citations. The headline rankings put La Roche-Posay first, with a Z-score of 86.8 and a 6.0% share of voice across the tested prompts, CeraVe second (Z-score 81.2, 5.0% share of voice), and Olaplex third (Z-score 71.7, 5.0% share of voice). Both La Roche-Posay and CeraVe appeared in 16.0% of all prompts analysed.
The payload underneath the ranking is what matters. Per the Avenue Z report, 63% of citations in the beauty category come from editorial sources. The specific publishers doing the citing are Allure.com (479 citations), Vogue.com (330), and Glamour.com (225). Reddit, which the general top-source chart would predict should dominate, contributed 153 citations. It is a serious signal, and it is third behind two glossy magazines.
What the three winners share, read from the citation data rather than the brand’s own marketing, is sustained coverage in a short list of editorial titles that AI engines treat as beauty authorities. The transferable lesson is therefore not that beauty brands should focus on Allure specifically. It is that in beauty the hidden citation layer is editorial, and nothing in the ChatGPT top-100 ranking tells you that.
In Supplements, the Model Prefers the Science Over the Revenue Leader
5W Research published the Supplements AI Visibility Index 2026 on 7 July 2026. It tested 3,800 prompts between January and April 2026 across ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews, covering the 25 most-cited supplement brands.
The headline is the gap between revenue and citation. AG1, widely reported as the revenue leader in the category, ranked third in observed AI citation frequency, behind Thorne and Seed. Four brands (Thorne, Seed, AG1 and Momentous) together captured an estimated more than 47% of observed citation share. 5W describes the gap between revenue rank and citation rank in supplements as wider than in any other sector they have measured.
Why Thorne wins where AG1 does not, in 5W’s own reading, is structural. Clinically cited product pages. Published research partnerships. Structured data on every product. Consistent mentions in health authority publications and clinical podcasts. The citation layer here is not editorial titles the way beauty is. It is medical publishers, peer-reviewed references, and the long tail of health podcasts that AI engines appear to be treating as expertise signals.
The transferable lesson for anyone running a store in a credentialled category, pharmacy, nutraceuticals, medical devices, legal, financial advice, is that revenue leadership and AI visibility are not the same race. The engines are weighting evidence over market presence, and the brand that invests in its evidence surface, whether or not it leads the category commercially, is the one being recommended at the top of the funnel.
In B2B Software, Review Sites Are Back as the Confidence Layer
The third report is G2’s 2026 Buyer Behavior Report, based on a March 2026 survey of 1,076 B2B software buyers and decision-makers. Its headline figure is that 51% of B2B software buyers now begin research with an AI chatbot more often than with Google, up from 29% a year earlier, and that 69% ended up choosing a different vendor from the one they initially planned, on AI chatbot guidance.
The relevant finding sits underneath that. In the same study, review sites rose to 38% as the top source shaping which vendors make a buyer’s shortlist, surpassing AI chatbots themselves at 37%. The chatbot is where discovery starts. The review site is still where the shortlist is formed. That is one direction in which the “AI replaces everything” narrative has not held up.
And underneath that is the vertical point. Paxton Gray, CEO of 97th Floor, made it directly in the WordPress VIP and 97th Floor “Building an AI-Ready Discoverability Strategy” webinar on 22 September 2026: for one of his B2B clients, the third most-cited source feeding AI answers was JazzHR, a vertical directory specific to recruiting software. Not Reddit. Not LinkedIn. Not G2 or Capterra as a horizontal. A vertical directory that most of his client’s marketing team had probably not actively engaged with in years.
The transferable lesson in B2B, then, has two layers. The horizontal review sites are the confidence check. The vertical directory of your specific category is where the model is being taught, and almost nobody is auditing their presence on those.
The Three Reports Agree on One Specific Thing
Read the three together and the agreement is tight. In beauty, editorial authority runs the citation layer. In supplements, clinical and research publishers run it. In B2B software, review sites and vertical directories run it. The actual sources are different, and that is the point. The disagreement with the generic “top-sources” chart is that there is no universal recipe. Each category has its own citation layer, and the first order of business is to figure out what yours looks like before pointing a budget at any of it.
What the three reports also share, more quietly, is that the citation layer is third-party by default. In every one of them, the sources teaching models about brands are not the brand’s own site. Your product pages matter, which is the next section. But they are evidence, not spine. The spine runs through publishers your category reads, review platforms your buyers trust, and directories or marketplaces that index your category.
What This Becomes Inside a Real WordPress or WooCommerce Install
The three reports describe where the models learn. The agency’s job is what you do about it, which splits into four work streams.
The first stream is on your own property, and it is relatively cheap: get the product and article schema right, every variant, every page, as JSON-LD baked into the template rather than left to an editor to remember. Core WooCommerce ships a Product schema that passes a validator and does nothing more. The useful work is extending it with the fields a model needs to attribute you correctly: brand, GTIN, SKU, aggregate rating, review count linked to actual reviews, and, where regulated, structured ingredient or specification data. A properly instrumented product template is roughly two to three days of build on an existing WooCommerce store, running into real cost only where custom fields need a new content model.
The second stream is the content model. The reason Thorne gets cited where AG1 does not is that the evidence lives in fields a model can read, not in paragraphs an editor wrote. If your clinical references, your award citations, your certifications, your partner names are inside free-text blocks on the page, they are illegible to the engines even when a human can see them. Fields with controlled vocabularies, exposed through the REST API or the Store API, are the fix. This is a one-off block editor and ACF or similar refactor per product line, and the uncomfortable part is that it almost always surfaces a decade of inconsistent terminology that nobody wants to normalise.
The third stream is off-site, and it is where the money lives. In beauty, that means PR relationships with the three editorial titles your category actually gets cited through, measured by citation count, not by vanity reach. In supplements, it means paid or earned placement in the clinical and medical publishers you now know matter. In B2B, it means paying attention to the one or two vertical directories specific to your category that AI engines treat as validation, and making sure your presence on them is complete, categorised, and current. The direct equivalent of claiming your Google Business Profile ten years ago, except on a different set of surfaces.
The fourth stream is measurement, and here I would be honest with the board. Current AI visibility tools, including the ones we run on client estates (Semrush AI Visibility, Profound, Peec AI, our own server log parsing), as Jodi Cerretani, CMO of WordPress VIP, has said on the record, measure what browsers show, which is humans. AI crawlers do not execute JavaScript, so attribution for them is already a different pipeline and is still partial. Server log analysis is where the real signal lives, and you should budget for the engineering it takes to make that readable monthly rather than quarterly.
In rough order of effort on a mid-market WordPress or WooCommerce estate, the four streams come out at about one month of implementation and roughly €18,000 to €35,000 for the first pass, depending on how clean the current product data is. The ongoing monthly cost is small on tools and larger on category-specific PR, which is where it should be: the citation layer that teaches the model lives outside your site.
If you are evaluating this work in-house or through a partner, the question I would ask before the first invoice is whether they can tell you which three to five sources your specific category gets cited through today. Not what the industry charts say. What yours shows. We run that analysis as a baseline on every new engagement, and more often than not it rewrites the plan rather than confirming it. If you want us to run yours, write to us, and we can go through the first pass together.
Last modified: October 16, 2026
